Dental caries identification method based on computer vision
Through the computer vision-based caries recognition method, using probability neural networks and physical prior embedding optimization strategies, the subjectivity and image feature complexity of traditional caries diagnosis are solved, and efficient and stable caries recognition is achieved.
Patent Information
- Application Number
- CN202510540929.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-04-27
- Publication Date
- 2025-08-08
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
Traditional dental caries diagnosis relies on manual interpretation and is susceptible to subjective factors. The image features are subtle and the density distribution is complex. Conventional image enhancement and classification methods are highly limited.
A computer vision-based caries recognition method is adopted to obtain image data for annotation and preprocessing, a probability neural network model is constructed, and a physical prior embedding and weight matrix optimization strategy is used to train and identify caries.
It improves the accuracy and stability of caries recognition, reduces the dependence on large-scale annotated data, enhances the performance of the model in unbalanced environments of noise and image quality, and is suitable for data-constrained scenarios.
Smart Images
Figure CN120451097A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of dental caries identification, and in particular relates to a dental caries identification method based on computer vision. Background Art
[0002] In the field of medical image analysis, the automated identification and grading of dental caries has long been an important area of research at the intersection of oral medicine and artificial intelligence. Traditional caries diagnosis relies primarily on dentists manually interpreting X-ray images (such as dental X-rays and CBCT images). This method not only relies on the physician's experience but is also susceptible to subjective factors in large-scale screening or diagnostic tasks, leading to inconsistent diagnostic results. Furthermore, the imaging features of dental caries are often subtle, and the density distribution of lesion areas varies complexly. Conventional image enhancement and classification methods have certain limitations when processing such medical images. Summary of the Invention
[0003] The purpose of the present invention is to provide a caries identification method based on computer vision, aiming to solve the problem that the image features of caries are often subtle, the density distribution of the lesion area varies complexly, and conventional image enhancement and classification methods have certain limitations when processing such medical images.
[0004] The present invention is achieved by providing a method for identifying dental caries based on computer vision, the method comprising:
[0005] Obtaining original dental caries image data, and labeling and preprocessing the original dental caries image data;
[0006] Construct a probabilistic neural network model and train it based on the preprocessed caries image data;
[0007] After the training is completed, the trained probabilistic neural network model is used to process new image data to identify caries contained therein.
[0008] Preferably, the step of preprocessing the original caries image data includes:
[0009] Size standardization: all original caries image data are uniformly adjusted to the same resolution;
[0010] Pixel normalization: A normalization strategy is used to map all pixel values to the range of 0, 1, which is expressed as:
[0011]
[0012] Where I(x,y) is the original pixel value, I norm (x, y) is the normalized pixel value, I min and I maxare the minimum and maximum pixel values of the image data respectively;
[0013] Denoising: Adaptive filtering is used to remove noise;
[0014] Contrast enhancement: Adaptive histogram equalization is used to improve contrast.
[0015] Preferably, the step of preprocessing the original caries image data further includes feature extraction:
[0016] Perform local feature extraction and use edge detection to extract detailed information of caries lesions, which can be expressed as:
[0017] G x =I enhanced *S x
[0018] G y =I enhanced *S y
[0019]
[0020] Where S x and S y are the filter kernels of the Sobel operator in the X and Y directions, I enhanced is the image after contrast enhancement, G x and G y is the gradient image, and G is the image after local feature extraction.
[0021] Preferably, the steps of constructing a probabilistic neural network model and training it based on the preprocessed caries image data include:
[0022] Build a probabilistic neural network model and initialize the weights and bias parameters of the probabilistic neural network model;
[0023] Adopting the inter-layer weight coupling synchronization strategy to monitor the correlation between weights of different layers, dynamically adjust the weight update method, and maintain information synchronization between network layers;
[0024] By optimizing the eigenvalues of the weight matrix, the weight change increment is calculated in each iteration to promote network convergence;
[0025] Cross entropy loss is used to evaluate the difference between the prediction and the true label;
[0026] Construct a constraint function based on the characteristics of tooth tissue to make the model learning process conform to the physical characteristics of tooth anatomy and optimize feature representation;
[0027] Repeat the iteration until the preset stopping condition is met.
[0028] Preferably, in the step of initializing the weight and bias parameters of the probabilistic neural network model, the weight W of the probabilistic neural network model is u and bias b u The initialization method is expressed as:
[0029]
[0030] Where, F u is the number of features after feature dimensionality reduction, C u is the number of categories, and rand() is an initialization function that randomly selects values from a uniform distribution.
[0031] Preferably, in the step of adopting the inter-layer weight coupling synchronization strategy to monitor the correlation between weights of different layers, dynamically adjust the weight update method, and maintain information synchronization between network layers, the weight update method of the probabilistic neural network model is expressed as:
[0032]
[0033]
[0034] In the formula, ΔW u is the weight change increment; sync() is the weight difference calculation function, which is used to calculate the weight difference between adjacent layers of the probabilistic neural network and use it to optimize weight updates; is the weight of the probabilistic neural network at the current t-th iteration, Lu is the number of layers of the probabilistic neural network, Denotes the loss function L u about The gradient, is the weight of the probabilistic neural network at the t+1th iteration, is the weight of the probabilistic neural network of the t-th iteration of the l-th layer, is the weight of the probabilistic neural network of the tth iteration of the l+1th layer, σ Ul () is the scaling function of the lth layer, η u is the learning rate, γ u is a coupling strength parameter that is a function of the difference in weights between adjacent layers.
[0035] Preferably, in the step of optimizing the eigenvalues of the weight matrix and calculating the weight change increment in each iteration to promote network convergence, the weight change increment is expressed as:
[0036]
[0037] In the formula, max() is the maximum value function; Calculate the eigenvalue of the weight matrix. The specific calculation method is to calculate the eigenvalue of the weight matrix Perform eigendecomposition and extract the values on the diagonal; α u is the eigenvalue adjustment rate; represents a normal distribution with a mean of 0 and a standard deviation of the identity matrix; is the noise intensity.
[0038] Preferably, in the step of using cross entropy loss to evaluate the difference between the prediction and the true label, the loss function is expressed as:
[0039]
[0040] Where, L u is the loss function, y u,i is the true label of the i-th sample, y u,i is the model prediction for the i-th sample, N u is the number of samples, and Lr is the physical prior embedding loss term.
[0041] Preferably, the constraint function based on the tooth tissue characteristics is constructed so that the learning process of the model conforms to the physical characteristics of the tooth anatomical structure, and in the step of optimizing the feature representation,
[0042]
[0043] Where, L r is the physical prior embedding loss term, N u is the number of pixels contained in the training sample, and Represent the dentin density value predicted by the model and the actual dentin density value, C s is the tooth structure consistency constraint, M u is the total number of tooth boundary pixels in the image, and Represent the predicted tooth boundary structure and the real boundary structure, Represents the gradient operator, which is used to measure the rate of change of the boundary. λ1 and λ2 are weight hyperparameters of the loss term, which are used to balance the density distribution error and the boundary gradient error. Preferably, λ1 is set to 0.5 and λ2 is set to 0.3.
[0044] Preferably, after the training is completed, the trained probabilistic neural network model is used to process the new image data, and in the step of identifying the caries contained therein, the new image data is preprocessed, including dimensioning, pixel normalization, denoising and contrast enhancement.
[0045] The computer vision-based caries recognition method provided by the present invention improves the model's ability to recognize caries categories through physical prior embedding and weight matrix optimization strategy, and improves the stability of the model under different image quality and noise environments through weight coupling synchronization mechanism and adaptive noise injection strategy, making the classification results more reliable. Physical prior knowledge enables the model to maintain high classification performance under small sample conditions, reduces dependence on large-scale labeled data, and facilitates the application of medical image analysis in data-limited scenarios. Through image standardization, denoising and contrast enhancement, the quality of input data is improved, enabling the classifier to more effectively learn the characteristics of caries lesions and reduce misjudgments caused by uneven image quality. BRIEF DESCRIPTION OF THE DRAWINGS
[0046] Figure 1 A diagram showing the performance comparison of different classifiers provided in an embodiment of the present invention;
[0047] Figure 2 A schematic diagram illustrating the impact of noise robustness on model stability provided by an embodiment of the present invention;
[0048] Figure 3 A schematic diagram showing the effect of data size on model performance provided by an embodiment of the present invention;
[0049] Figure 4 A schematic diagram illustrating the effect of preprocessing on classification accuracy provided by an embodiment of the present invention. DETAILED DESCRIPTION
[0050] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.
[0051] An embodiment of the present invention provides a method for identifying dental caries based on computer vision, the method comprising:
[0052] Obtaining original dental caries image data, and labeling and preprocessing the original dental caries image data;
[0053] Construct a probabilistic neural network model and train it based on the preprocessed caries image data;
[0054] After the training is completed, the trained probabilistic neural network model is used to process new image data to identify caries contained therein.
[0055] In an embodiment of the present invention, the computer vision-based dental caries identification method comprises the following steps:
[0056] S1. Collect, label and preprocess the original caries image data.
[0057] The specific process of data input is as follows:
[0058] S101. Caries image data collection:
[0059] Dental X-ray images (including dental X-rays, cone beam CT (CBCT) images, etc.) are used as data sources. The image data comes from:
[0060] 1) Public datasets: such as Aphrodite, FIDTM and other medical imaging datasets.
[0061] 2) Clinical acquisition: Use dental X-ray equipment that meets medical standards to obtain patient dental images, ensuring that the image resolution and contrast are suitable for medical diagnosis.
[0062] When collecting data, it is necessary to ensure that all images are acquired under uniform conditions to avoid the impact of factors such as lighting and equipment parameters on image quality.
[0063] S102. Annotation of image data:
[0064] In order to ensure the validity of the training data, caries images were annotated and classified into different categories (such as shallow caries, moderate caries, and deep caries) according to the severity of caries lesions.
[0065] S103, image data preprocessing:
[0066] Before inputting into the classifier, the image needs to be standardized to reduce the impact of noise and improve classification accuracy. This mainly includes:
[0067] 1) Size standardization:
[0068] Since the image sizes of different data sources may be different, all images are resized to the same resolution (H, W), such as 256×256 or 512×512.
[0069] 2) Pixel normalization:
[0070] Since the grayscale value range of medical images is relatively wide (usually pixel intensity of 0 to 255), a normalization strategy is adopted to map all pixel values to the range of 0, 1, which can be expressed as:
[0071]
[0072] Where I(x,y) is the original pixel value, I norm (x, y) is the normalized pixel value, I min and I max are the minimum and maximum pixel values of the image data, respectively.
[0073] 3) Denoising:
[0074] Since there may be noise in X-ray images, an adaptive filtering method is used to remove irrelevant noise, which can be expressed as:
[0075]
[0076] Where G(m,n) is a two-dimensional Gaussian kernel, k is the filter window size, and I filtered (x,y) is the pixel value after filtering.
[0077] 4) Contrast enhancement:
[0078] In order to improve the visibility of the caries area, the adaptive histogram equalization method is used to improve the image contrast, which is expressed as:
[0079]
[0080] Where, I local (x, y) is the pixel value in the local window after denoising, I mean and I std are the mean and standard deviation of the local window, I enhanced (x,y) is the pixel value after contrast enhancement.
[0081] S104, feature extraction:
[0082] After the image preprocessing is completed, the key features of the image are obtained so that the classifier can learn. The present invention adopts a multi-scale feature extraction strategy to improve the recognition ability of the classifier by fusing local features with global features.
[0083] First, local feature extraction is performed, and edge detection (such as the Sobel operator) is used to extract the detail information of the caries lesion, which is expressed as:
[0084] G x =I enhanced *S x
[0085] G y =I enhanced *S y
[0086]
[0087] Where S x and S y are the filter kernels of the Sobel operator in the X and Y directions, I enhanced is the image after contrast enhancement, G x and G y is the gradient image, and G is the image after local feature extraction.
[0088] S2. After data preprocessing, the caries image data is input into the classifier for training.
[0089] A probabilistic neural network is used as the classifier model to classify project risk categories. Traditional probabilistic neural networks may encounter category imbalance and feature dependency problems when processing classification tasks. The weight matrix eigenvalue optimization strategy is used to adjust the weight coupling between network layers and optimize the decision boundary of the classifier to make it more suitable for different risk assessment scenarios.
[0090] Specifically, the training process of the probabilistic neural network algorithm based on weight matrix eigenvalue optimization is as follows:
[0091] S201, initialize the weight and bias parameters of the probabilistic neural network. In one embodiment, the weight W of the probabilistic neural network u and bias b u The initialization method is expressed as:
[0092]
[0093] Where, F u is the number of features after feature dimensionality reduction, C u is the number of categories, and rand() is an initialization function that randomly selects values from a uniform distribution.
[0094] S202. During the network training process, an inter-layer weight coupling synchronization strategy is used to monitor the correlation between weights of different layers and dynamically adjust the weight update method to maintain information synchronization between network layers. The weight update method of the probabilistic neural network is expressed as:
[0095]
[0096]
[0097] In the formula, ΔW u is the weight change increment; sync() is the weight difference calculation function, which is used to calculate the weight difference between adjacent layers of the probabilistic neural network and use it to optimize weight updates; is the weight of the probabilistic neural network at the current t-th iteration, Lu is the number of layers of the probabilistic neural network, Denotes the loss function L u about The gradient, is the weight of the probabilistic neural network at the t+1th iteration, is the weight of the probabilistic neural network of the t-th iteration of the l-th layer, is the weight of the probabilistic neural network of the tth iteration of the l+1th layer, σ Ul() is the scaling function of the lth layer, η u is the learning rate, γ u is a coupling strength parameter that is a function of the difference in weights of adjacent layers. u Set to 0.3, L u Set to 5.
[0098] Scaling function σ Ul Used to balance the contribution of weight differences between different layers, depending on the size of the layer or the complexity of the connection between layers, with weights For example, The calculation method is expressed as:
[0099]
[0100] Where, is the Frobenius norm, Represents the total number of elements in the matrix.
[0101] S203, using the eigenvalues of the weight matrix for optimization, in each iteration process, the weight change increment is calculated to make the network converge to the optimized state faster. The calculation method is expressed as:
[0102]
[0103] In the formula, max() is the maximum value function; Calculate the eigenvalue of the weight matrix. The specific calculation method is to calculate the eigenvalue of the weight matrix Perform eigendecomposition and then extract the values on its diagonal; α u is the eigenvalue adjustment rate; represents a normal distribution with a mean of 0 and a standard deviation of the identity matrix; is the noise intensity. Preferably, α u Set to 0.3.
[0104] In one embodiment, the noise intensity adopts an adaptive noise injection method to make the network show better stability and adaptability when facing possible data disturbances. In each training iteration, the loss function is dynamically adjusted according to its rate of change. The adjustment method is expressed as:
[0105]
[0106] Where K u is the baseline value of the noise intensity, δ u is the adjustment factor, is the derivative of the loss function with respect to the number of iterations. Preferably, K u Set to 3, δ u Set to 2.
[0107] S204. During the training process, the cross entropy loss is used to evaluate the difference between the prediction and the true label. The calculation method is expressed as:
[0108]
[0109] Where, L u is the loss function, y u,i is the true label of the i-th sample, is the model prediction for the i-th sample, N u is the number of samples, and Lr is the physical prior embedding loss term.
[0110] S206, the physical prior embedding loss term constrains the model to learn features that conform to medical laws by considering prior knowledge of tooth anatomy (such as enamel / dentin density distribution), thereby improving classification accuracy and generalization ability. The present invention constructs a constraint function based on tooth tissue characteristics to make the model learning process conform to the physical characteristics of tooth anatomy, optimize feature representation, and improve classification credibility. The calculation method is expressed as:
[0111]
[0112] Where, L r is the physical prior embedding loss term, N u is the number of pixels contained in the training sample, and Represent the dentin density value predicted by the model and the actual dentin density value, C s is the tooth structure consistency constraint, M u is the total number of tooth boundary pixels in the image, and Represent the predicted tooth boundary structure and the real boundary structure, Represents the gradient operator, which is used to measure the rate of change of the boundary. λ1 and λ2 are weight hyperparameters of the loss term, which are used to balance the density distribution error and the boundary gradient error. Preferably, λ1 is set to 0.5 and λ2 is set to 0.3.
[0113] In this embodiment of the present invention, a strategy of tooth structure consistency constraint is used to impose smooth constraints on the density gradients of adjacent pixels during the network learning process, so that the model can better learn the density changes between enamel and dentin:
[0114]
[0115] Where C s is the tooth structure consistency constraint, and Represent the density gradients in the x and y directions respectively, pred and true represent the model prediction value and the true value respectively.
[0116] S207, repeat the above steps until the preset stop iteration condition is met, which means that the model training is completed. In one embodiment, the preset stop iteration condition is reaching a preset maximum number of iterations. Preferably, the preset maximum number of iterations is set to 1000 times.
[0117] S3. Caries Identification:
[0118] After completing the model training, the trained probabilistic neural network model is used to identify caries on new data.
[0119] The data to be detected is input into the trained model for recognition. To ensure the recognition effect, the input image needs to be preprocessed in the same way as the training data, including:
[0120] Size normalization: adjust the image to a uniform resolution to ensure consistency of model input;
[0121] Pixel normalization: Normalizes image pixel values to a specific range to match the format of training data;
[0122] Denoising: Adaptive filtering is used to remove noise and improve model robustness;
[0123] Contrast Enhancement: Uses adaptive histogram equalization to enhance image contrast and improve the visibility of caries areas.
[0124] In an embodiment of the present invention, multi-scale features of the image are extracted through a trained model, including local features (such as edges and gradient information) and global features (such as density change trends in caries areas), and the Sobel operator is used to extract caries edge information to enhance the features of the caries lesion area.
[0125] In the embodiment of the present invention, a trained probabilistic neural network model is used to classify the input image and identify the type of dental caries (shallow caries, moderate caries, deep caries).
[0126] Experimental verification:
[0127] like Figure 1As shown in the figure, a performance comparison study of different classifiers was conducted to verify the effectiveness of the improved probabilistic neural network in the caries grading task. The experiment selected traditional convolutional neural networks, support vector machines, and random forests as comparison objects. By measuring classification accuracy and comprehensive evaluation indicators, the ability of each method to identify caries lesion characteristics was systematically evaluated. The experimental results show that the improved probabilistic neural network significantly outperformed other methods in both indicators. This is due to the constraints on tooth density distribution by physical prior knowledge and the improvement of the decision boundary brought about by the optimization of the eigenvalues of the weight matrix, which enables the model to more accurately distinguish subtle differences in caries lesions.
[0128] like Figure 2 As shown in the figure, to explore the impact of noise interference on model stability, the noise resistance capabilities of traditional convolutional neural networks and the proposed method are compared. By simulating noise pollution scenarios of varying intensities, the performance of the model is observed as the noise increases. Experiments show that the performance of the traditional method decreases sharply as the noise intensity increases. However, the proposed method, through an adaptive noise injection strategy and a weight coupling synchronization mechanism, effectively alleviates the feature distortion caused by noise interference, verifying the robustness advantage of the eigenvalue optimization strategy in complex medical imaging environments.
[0129] like Figure 3 As shown in the figure, by changing the scale of training data, the role of physical prior embedding in promoting small sample learning is verified. Unlike traditional methods that rely on massive data, the experimental results show that under the conditions of limited training samples, the method of the present invention can still maintain a high accuracy rate, indicating that the additional constraint information provided by the prior knowledge of dental anatomy enables the model to extract essential features that conform to medical laws from limited data, significantly reducing the dependence on the amount of data.
[0130] like Figure 4 To validate the necessity of a multi-stage preprocessing pipeline, the authors gradually superimposed normalization, denoising, and contrast enhancement steps to examine the impact of each on the final classification. The experiments revealed that the complete preprocessing pipeline significantly improved performance compared to the original data, with adaptive filtering to remove device noise and histogram equalization to enhance lesion contrast contributing the most. This demonstrates that the image quality optimization and feature enhancement strategies effectively highlight the morphological characteristics of caries lesions, providing a high-quality input data foundation for subsequent classifiers.
[0131] The above-described embodiments merely illustrate several implementations of the present invention, and while their descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that a person skilled in the art would be able to make numerous variations and improvements without departing from the spirit of the present invention, all of which fall within the scope of protection of the present invention. Therefore, the scope of protection of the present invention shall be determined by the appended claims.
[0132] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.
Claims
1. A method for identifying dental caries based on computer vision, characterized in that: The method comprises: Obtaining original dental caries image data, and labeling and preprocessing the original dental caries image data; Construct a probabilistic neural network model and train it based on the preprocessed caries image data; After the training is completed, the trained probabilistic neural network model is used to process new image data to identify caries contained therein.
2. The computer vision-based caries identification method according to claim 1, characterized in that: The steps of preprocessing the original caries image data include: Size standardization: all original caries image data are uniformly adjusted to the same resolution; Pixel normalization: A normalization strategy is used to map all pixel values to the range of 0, 1, which is expressed as: Where I(x,y) is the original pixel value, I norm (x, y) is the normalized pixel value, I min and I max are the minimum and maximum pixel values of the image data respectively; Denoising: Adaptive filtering is used to remove noise; Contrast enhancement: Adaptive histogram equalization is used to improve contrast.
3. The computer vision-based caries identification method according to claim 2, characterized in that: The steps of preprocessing the original caries image data also include feature extraction: Perform local feature extraction and use edge detection to extract detailed information of caries lesions, which can be expressed as: G x =I enhanced *S x G y =I enhanced *S y Where S x and S y are the filter kernels of the Sobel operator in the X and Y directions, I enhanced is the image after contrast enhancement, G x and G y is the gradient image, and G is the image after local feature extraction.
4. The computer vision-based caries identification method according to claim 1, characterized in that: The steps of constructing a probabilistic neural network model and training it based on the preprocessed caries image data include: Build a probabilistic neural network model and initialize the weights and bias parameters of the probabilistic neural network model; Adopting the inter-layer weight coupling synchronization strategy to monitor the correlation between weights of different layers, dynamically adjust the weight update method, and maintain information synchronization between network layers; By optimizing the eigenvalues of the weight matrix, the weight change increment is calculated in each iteration to promote network convergence; Cross entropy loss is used to evaluate the difference between the prediction and the true label; Construct a constraint function based on the characteristics of tooth tissue to make the model learning process conform to the physical characteristics of tooth anatomy and optimize feature representation; Repeat the iteration until the preset stopping condition is met.
5. The computer vision-based caries identification method according to claim 4, characterized in that: In the step of initializing the weight and bias parameters of the probabilistic neural network model, the weight W of the probabilistic neural network model is u and bias b u The initialization method is expressed as: Where, F u is the number of features after feature dimensionality reduction, C u is the number of categories, and rand() is an initialization function that randomly selects values from a uniform distribution.
6. The computer vision-based caries identification method according to claim 4, characterized in that: In the step of adopting the inter-layer weight coupling synchronization strategy to monitor the correlation between weights of different layers, dynamically adjust the weight update method, and maintain information synchronization between network layers, the weight update method of the probabilistic neural network model is expressed as: In the formula, ΔW u is the weight change increment; sync() is the weight difference calculation function, which is used to calculate the weight difference between adjacent layers of the probabilistic neural network and use it to optimize weight updates; is the weight of the probabilistic neural network at the current t-th iteration, Lu is the number of layers of the probabilistic neural network, Denotes the loss function L u about The gradient, is the weight of the probabilistic neural network at the t+1th iteration, is the weight of the probabilistic neural network of the t-th iteration of the l-th layer, is the weight of the probabilistic neural network of the tth iteration of the l+1th layer, σ Ul () is the scaling function of the lth layer, η u is the learning rate, γ u is a coupling strength parameter that is a function of the difference in weights between adjacent layers.
7. The computer vision-based caries identification method according to claim 4, characterized in that: In the step of optimizing the eigenvalues of the weight matrix and calculating the weight change increment in each iteration to promote network convergence, the weight change increment is expressed as: In the formula, max() is the maximum value function; Calculate the eigenvalue of the weight matrix. The specific calculation method is to calculate the eigenvalue of the weight matrix Perform eigendecomposition and extract the values on the diagonal; α u is the eigenvalue adjustment rate; represents a normal distribution with a mean of 0 and a standard deviation of the identity matrix; is the noise intensity.
8. The computer vision-based caries identification method according to claim 4, characterized in that: In the step of using cross entropy loss to evaluate the difference between the prediction and the true label, the loss function is expressed as: Where, L u is the loss function, y u,i is the true label of the i-th sample, is the model prediction for the i-th sample, N u is the number of samples, and Lr is the physical prior embedding loss term.
9. The computer vision-based caries identification method according to claim 4, characterized in that: The constraint function based on the tooth tissue characteristics is constructed to make the learning process of the model conform to the physical characteristics of the tooth anatomical structure, and in the step of optimizing the feature representation, Where, L r is the physical prior embedding loss term, N u is the number of pixels contained in the training sample, and Represent the dentin density value predicted by the model and the actual dentin density value, C s is the tooth structure consistency constraint, M u is the total number of tooth boundary pixels in the image, and Represent the predicted tooth boundary structure and the real boundary structure, Represents the gradient operator, which is used to measure the rate of change of the boundary. λ1 and λ2 are weight hyperparameters of the loss term, which are used to balance the density distribution error and the boundary gradient error. Preferably, λ1 is set to 0.5 and λ2 is set to 0.
3.
10. The computer vision-based caries identification method according to claim 1, characterized in that: After the training is completed, the trained probabilistic neural network model is used to process the new image data. In the step of identifying the caries contained therein, the new image data is preprocessed, including dimensioning, pixel normalization, denoising and contrast enhancement.